液晶与显示, 2019, 34 (1): 110, 网络出版: 2019-03-06   

角度空间三元组损失微调的人脸识别

Face recognition of triple loss fine-tuning in angular space
作者单位
江西理工大学 信息工程学院, 江西 赣州 341000
摘要
针对角度Softmax损失强约束存在的问题, 提出一种用角度空间三元组损失对角度Softmax损失预训练模型进行微调的算法。算法首先对原来的卷积神经网络结构进行改进, 将1×1卷积核与池化层加在不同残差块间, 以进行选择更有效的特征。然后用角度空间下的三元组损失对预训练模型进行微调, 以降低困难样本的强约束条件。最后在测试时, 分别提取原始人脸图像特征和水平翻转的人脸图像特征, 对两个特征相加作为最终的人脸特征表达, 以丰富人脸特征信息, 提高识别率。实验结果表明, 在LFW和YTF人脸数据集分别取得了99.25%和94.52%的识别率, 在大规模人脸身份识别中, 本文提出的方法在仅用单模型和比较小的训练集就能有效地提高人脸识别率。
Abstract
Aiming at the problem that angular Softmax loss exists in strong constraint, an algorithm to fine-turn the angular Softmax loss pre-training model by using triple loss in angle space was proposed in this paper. Firstly, the original structure of convolution neural network was improved and the convolution kernels of 1×1 and pooling layers were added between different residual blocks to select more effective features. Secondly, the pre-training model was fine-tuned by triple loss in angular space to reduce the strong constraint condition of difficult samples. Finally, in the test, the features of original and the horizontal flipped face images were extracted respectively, and the two features were added as the final facial feature expression, so as to enrich the face feature information and improve the recognition rate. The experimental results show that the recognition rates of the LFW and YTF face data sets are 99.25% and 94.52%, respectively. The method proposed in this paper can effectively improve the face recognition rate in large-scale face identification by using only a single model and a relatively small training set.

任克强, 胡慧. 角度空间三元组损失微调的人脸识别[J]. 液晶与显示, 2019, 34(1): 110. REN Ke-qiang, HU Hui. Face recognition of triple loss fine-tuning in angular space[J]. Chinese Journal of Liquid Crystals and Displays, 2019, 34(1): 110.

本文已被 1 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!